hyperhive/hive-ag3nt/src/turn.rs
müde 6db38cf70c model: runtime override via /model slash; fixes for port + bind
- runtime model override: Bus::{model,set_model} + POST /api/model
  (form-encoded {model: name}). turn.rs reads bus.model() per turn
  so a flip lands on the next claude invocation. /api/state grows
  a model field; agent page shows a 'model · <name>' chip in the
  state row. '/model <name>' slash command POSTs to the endpoint
  and refreshes state.

- port regression fix: agent_web_port no longer probes forward for
  *existing* agents (the previous fix shifted ports for any agent
  without a port file, including legacy ones whose container was
  already bound to the bare hashed port — dashboard rendered the
  new port, container was still on the old one, conn errors). new
  rule: port file exists → use it; absent + applied flake present
  → legacy, persist port_hash without probing; absent + no applied
  flake → fresh spawn, probe forward.

- SO_REUSEADDR on both the dashboard and per-agent web UI binds
  via tokio::net::TcpSocket. operator hit 12 retries failing on
  manager :8000 — REUSEADDR handles the TIME_WAIT case cleanly
  without a new dep; retry still covers the genuine
  process-still-alive overlap.

todo: drops the model-override entry (shipped); adds two new
items — model persistence (optional, future), and custom
per-agent MCP tools (groundwork for moving bitburner-agent into
hyperhive).
2026-05-15 20:59:45 +02:00

302 lines
11 KiB
Rust

//! Per-turn claude invocation shared by `hive-ag3nt` and `hive-m1nd`. The
//! two binaries differ only in their MCP `Flavor` (agent surface vs.
//! manager surface) and their wake-prompt wording; the spawn shape,
//! arg-vector, stdin plumbing, and stream-json pumping are identical.
use std::path::{Path, PathBuf};
use std::process::Stdio;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::{Arc, Mutex};
use std::time::Duration;
use anyhow::{Result, bail};
use tokio::io::{AsyncBufReadExt, AsyncWriteExt, BufReader};
use tokio::process::Command;
use crate::events::{Bus, LiveEvent};
use crate::login::{self, LoginState};
use crate::mcp;
/// `--settings` JSON applied to every claude invocation. Lives as a
/// properly-formatted file in `prompts/claude-settings.json` so it's easy
/// to read and edit; we ship it via `include_str!`. We turn off claude's
/// in-session auto-compaction and its cross-session auto-memory because
/// hyperhive owns those concerns (`/compact` on overflow, notes
/// persistence under `/state`). Unknown keys are silently ignored by
/// claude-code; if a key gets renamed we'll spot it because the
/// corresponding behavior will start firing mid-turn again.
const CLAUDE_SETTINGS: &str = include_str!("../prompts/claude-settings.json");
/// Regex-ish marker claude-code emits when context overflows. Same string
/// bitburner-agent watches for. Empirically reliable across claude-code
/// versions; if it ever changes, compaction won't fire and we'll see a
/// claude exit with a useful error in the live view.
const PROMPT_TOO_LONG_MARKER: &str = "Prompt is too long";
/// Drop the MCP config blob claude reads from `--mcp-config <path>`.
/// `socket` is the hyperhive per-container socket (forwarded to the child
/// as `--socket <path>`); `binary_subcommand` is e.g. `"mcp"` for sub-agents
/// or `"mcp"` for the manager (both binaries name their MCP subcommand the
/// same — the differentiator is which binary `/proc/self/exe` resolves to).
pub async fn write_mcp_config(socket: &Path) -> Result<PathBuf> {
let parent = socket.parent().unwrap_or_else(|| Path::new("/run/hive"));
tokio::fs::create_dir_all(parent).await.ok();
let path = parent.join("claude-mcp-config.json");
let exe = std::env::current_exe()
.ok()
.map_or_else(|| "hive-ag3nt".into(), |p| p.display().to_string());
let body = mcp::render_claude_config(&exe, socket);
tokio::fs::write(&path, body).await?;
tracing::info!(path = %path.display(), "wrote claude MCP config");
Ok(path)
}
/// Drop the static `--settings` JSON next to the MCP config so we can
/// pass a path (`--settings <file>`) instead of an ever-growing inline
/// blob — the CLI argv has a finite length budget.
pub async fn write_settings(socket: &Path) -> Result<PathBuf> {
let parent = socket.parent().unwrap_or_else(|| Path::new("/run/hive"));
tokio::fs::create_dir_all(parent).await.ok();
let path = parent.join("claude-settings.json");
tokio::fs::write(&path, CLAUDE_SETTINGS).await?;
tracing::info!(path = %path.display(), "wrote claude settings");
Ok(path)
}
/// Write the agent's / manager's static system prompt to a file next to
/// the MCP config and return the path. Passed to claude via
/// `--system-prompt-file`, replacing claude's default system prompt with
/// the role + tools instructions. Per-turn prompts become much smaller
/// (just the wake message body).
pub async fn write_system_prompt(
socket: &Path,
label: &str,
flavor: mcp::Flavor,
) -> Result<PathBuf> {
let parent = socket.parent().unwrap_or_else(|| Path::new("/run/hive"));
tokio::fs::create_dir_all(parent).await.ok();
let template = match flavor {
mcp::Flavor::Agent => include_str!("../prompts/agent.md"),
mcp::Flavor::Manager => include_str!("../prompts/manager.md"),
};
let body = template.replace("{label}", label);
let path = parent.join("claude-system-prompt.md");
tokio::fs::write(&path, body).await?;
tracing::info!(path = %path.display(), "wrote claude system prompt");
Ok(path)
}
/// One claude turn's outcome. The harness uses this to decide whether to
/// transparently kick off a compaction and retry.
#[derive(Debug)]
pub enum TurnOutcome {
Ok,
/// claude saw "Prompt is too long" — the session needs compacting.
/// Run `compact_session()` then retry the same wake-up prompt.
PromptTooLong,
Failed(anyhow::Error),
}
/// Drive one turn end-to-end, transparently compacting + retrying once on
/// `Prompt is too long`. Both the sub-agent and manager loops call this.
pub async fn drive_turn(
prompt: &str,
mcp_config: &Path,
system_prompt: &Path,
settings: &Path,
bus: &Bus,
flavor: mcp::Flavor,
) -> TurnOutcome {
match run_turn(prompt, mcp_config, system_prompt, settings, bus, flavor).await {
TurnOutcome::PromptTooLong => {
if let Err(e) = compact_session(settings, bus).await {
tracing::warn!(error = %format!("{e:#}"), "compact failed");
return TurnOutcome::Failed(e);
}
run_turn(prompt, mcp_config, system_prompt, settings, bus, flavor).await
}
other => other,
}
}
/// Emit the per-turn `TurnEnd` event + log line. Single owner so the
/// agent and manager loops agree on outcome semantics.
pub fn emit_turn_end(bus: &Bus, outcome: &TurnOutcome) {
match outcome {
TurnOutcome::Ok | TurnOutcome::PromptTooLong => {
bus.emit(LiveEvent::TurnEnd {
ok: true,
note: None,
});
tracing::info!("turn finished");
}
TurnOutcome::Failed(e) => {
let note = format!("{e:#}");
bus.emit(LiveEvent::TurnEnd {
ok: false,
note: Some(note.clone()),
});
tracing::warn!(error = %note, "turn failed");
}
}
}
/// Block until the bound `~/.claude/` dir contains a session, polling
/// `claude_dir` on a `poll_ms` interval (min 2s). Flips `state` to
/// `Online` when login lands; caller resumes its serve loop.
pub async fn wait_for_login(claude_dir: &Path, state: Arc<Mutex<LoginState>>, poll_ms: u64) {
tracing::warn!(
claude_dir = %claude_dir.display(),
"no claude session — staying in partial-run mode (web UI only)"
);
let probe = Duration::from_millis(poll_ms.max(2000));
loop {
tokio::time::sleep(probe).await;
if login::has_session(claude_dir) {
tracing::info!("claude session detected — entering turn loop");
*state.lock().unwrap() = LoginState::Online;
return;
}
}
}
/// Spawn `claude` for one turn and pump `stream-json` stdout into the
/// live event bus. Prompt goes over stdin (variadic
/// `--allowedTools`/`--tools` would otherwise eat a trailing positional
/// prompt). The session is persistent across turns via `--continue` and
/// claude's in-session auto-compact is disabled via `--settings` so it
/// doesn't stall mid-turn — hyperhive owns compaction.
pub async fn run_turn(
prompt: &str,
mcp_config: &Path,
system_prompt: &Path,
settings: &Path,
bus: &Bus,
flavor: mcp::Flavor,
) -> TurnOutcome {
match run_claude(
prompt,
mcp_config,
Some(system_prompt),
settings,
bus,
flavor,
ClaudeMode::Turn,
)
.await
{
Ok(too_long) if too_long => TurnOutcome::PromptTooLong,
Ok(_) => TurnOutcome::Ok,
Err(e) => TurnOutcome::Failed(e),
}
}
/// Run claude's built-in `/compact` slash command on the persistent
/// session so the next turn can fit. No MCP tools needed; we just feed
/// `/compact` over stdin and let claude rewrite its own history.
pub async fn compact_session(settings: &Path, bus: &Bus) -> Result<()> {
bus.emit(LiveEvent::Note(
"context overflow — running /compact on the persistent session".into(),
));
let _ = run_claude(
"/compact",
Path::new("/dev/null"),
None,
settings,
bus,
mcp::Flavor::Agent, // tool surface unused for /compact
ClaudeMode::Compact,
)
.await?;
bus.emit(LiveEvent::Note("/compact done".into()));
Ok(())
}
#[derive(Clone, Copy)]
enum ClaudeMode {
Turn,
Compact,
}
async fn run_claude(
prompt: &str,
mcp_config: &Path,
system_prompt: Option<&Path>,
settings: &Path,
bus: &Bus,
flavor: mcp::Flavor,
mode: ClaudeMode,
) -> Result<bool> {
let model = bus.model();
let mut cmd = Command::new("claude");
cmd.arg("--print")
.arg("--verbose")
.arg("--output-format")
.arg("stream-json")
.arg("--model")
.arg(&model)
.arg("--continue")
.arg("--settings")
.arg(settings);
if let Some(p) = system_prompt {
cmd.arg("--system-prompt-file").arg(p);
}
if let ClaudeMode::Turn = mode {
cmd.arg("--mcp-config")
.arg(mcp_config)
.arg("--strict-mcp-config")
.arg("--tools")
.arg(mcp::builtin_tools_arg())
.arg("--allowedTools")
.arg(mcp::allowed_tools_arg(flavor));
}
let mut child = cmd
.stdin(Stdio::piped())
.stdout(Stdio::piped())
.stderr(Stdio::piped())
.spawn()?;
if let Some(mut stdin) = child.stdin.take() {
stdin.write_all(prompt.as_bytes()).await?;
stdin.shutdown().await.ok();
drop(stdin);
}
let stdout = child.stdout.take().expect("piped stdout");
let stderr = child.stderr.take().expect("piped stderr");
let prompt_too_long = Arc::new(AtomicBool::new(false));
let flag_out = prompt_too_long.clone();
let flag_err = prompt_too_long.clone();
let bus_out = bus.clone();
let bus_err = bus.clone();
let pump_stdout = tokio::spawn(async move {
let mut reader = BufReader::new(stdout).lines();
while let Ok(Some(line)) = reader.next_line().await {
if line.contains(PROMPT_TOO_LONG_MARKER) {
flag_out.store(true, Ordering::Relaxed);
}
match serde_json::from_str::<serde_json::Value>(&line) {
Ok(v) => bus_out.emit(LiveEvent::Stream(v)),
Err(_) => bus_out.emit(LiveEvent::Note(format!("(non-json) {line}"))),
}
}
});
let pump_stderr = tokio::spawn(async move {
let mut reader = BufReader::new(stderr).lines();
while let Ok(Some(line)) = reader.next_line().await {
if line.contains(PROMPT_TOO_LONG_MARKER) {
flag_err.store(true, Ordering::Relaxed);
}
bus_err.emit(LiveEvent::Note(format!("stderr: {line}")));
}
});
let status = child.wait().await?;
let _ = pump_stdout.await;
let _ = pump_stderr.await;
let too_long = prompt_too_long.load(Ordering::Relaxed);
if !status.success() && !too_long {
bail!("claude exited {status}");
}
Ok(too_long)
}